1 //===-- LoopIdiomRecognize.cpp - Loop idiom recognition -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass implements an idiom recognizer that transforms simple loops into a
11 // non-loop form. In cases that this kicks in, it can be a significant
14 //===----------------------------------------------------------------------===//
18 // Future loop memory idioms to recognize:
19 // memcmp, memmove, strlen, etc.
20 // Future floating point idioms to recognize in -ffast-math mode:
22 // Future integer operation idioms to recognize:
25 // Beware that isel's default lowering for ctpop is highly inefficient for
26 // i64 and larger types when i64 is legal and the value has few bits set. It
27 // would be good to enhance isel to emit a loop for ctpop in this case.
29 // We should enhance the memset/memcpy recognition to handle multiple stores in
30 // the loop. This would handle things like:
31 // void foo(_Complex float *P)
32 // for (i) { __real__(*P) = 0; __imag__(*P) = 0; }
33 // this is also "Example 2" from http://blog.regehr.org/archives/320
35 // This could regognize common matrix multiplies and dot product idioms and
36 // replace them with calls to BLAS (if linked in??).
38 //===----------------------------------------------------------------------===//
40 #define DEBUG_TYPE "loop-idiom"
41 #include "llvm/Transforms/Scalar.h"
42 #include "llvm/Analysis/AliasAnalysis.h"
43 #include "llvm/Analysis/LoopPass.h"
44 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
45 #include "llvm/Analysis/ScalarEvolutionExpander.h"
46 #include "llvm/Analysis/ValueTracking.h"
47 #include "llvm/Target/TargetData.h"
48 #include "llvm/Transforms/Utils/Local.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/IRBuilder.h"
51 #include "llvm/Support/raw_ostream.h"
52 #include "llvm/ADT/Statistic.h"
55 STATISTIC(NumMemSet, "Number of memset's formed from loop stores");
56 STATISTIC(NumMemCpy, "Number of memcpy's formed from loop load+stores");
59 class LoopIdiomRecognize : public LoopPass {
66 explicit LoopIdiomRecognize() : LoopPass(ID) {
67 initializeLoopIdiomRecognizePass(*PassRegistry::getPassRegistry());
70 bool runOnLoop(Loop *L, LPPassManager &LPM);
71 bool runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
72 SmallVectorImpl<BasicBlock*> &ExitBlocks);
74 bool processLoopStore(StoreInst *SI, const SCEV *BECount);
76 bool processLoopStoreOfSplatValue(StoreInst *SI, unsigned StoreSize,
78 const SCEVAddRecExpr *Ev,
80 bool processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
81 const SCEVAddRecExpr *StoreEv,
82 const SCEVAddRecExpr *LoadEv,
85 /// This transformation requires natural loop information & requires that
86 /// loop preheaders be inserted into the CFG.
88 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
89 AU.addRequired<LoopInfo>();
90 AU.addPreserved<LoopInfo>();
91 AU.addRequiredID(LoopSimplifyID);
92 AU.addPreservedID(LoopSimplifyID);
93 AU.addRequiredID(LCSSAID);
94 AU.addPreservedID(LCSSAID);
95 AU.addRequired<AliasAnalysis>();
96 AU.addPreserved<AliasAnalysis>();
97 AU.addRequired<ScalarEvolution>();
98 AU.addPreserved<ScalarEvolution>();
99 AU.addPreserved<DominatorTree>();
100 AU.addRequired<DominatorTree>();
105 char LoopIdiomRecognize::ID = 0;
106 INITIALIZE_PASS_BEGIN(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
108 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
109 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
110 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
111 INITIALIZE_PASS_DEPENDENCY(LCSSA)
112 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
113 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
114 INITIALIZE_PASS_END(LoopIdiomRecognize, "loop-idiom", "Recognize loop idioms",
117 Pass *llvm::createLoopIdiomPass() { return new LoopIdiomRecognize(); }
119 /// DeleteDeadInstruction - Delete this instruction. Before we do, go through
120 /// and zero out all the operands of this instruction. If any of them become
121 /// dead, delete them and the computation tree that feeds them.
123 static void DeleteDeadInstruction(Instruction *I, ScalarEvolution &SE) {
124 SmallVector<Instruction*, 32> NowDeadInsts;
126 NowDeadInsts.push_back(I);
128 // Before we touch this instruction, remove it from SE!
130 Instruction *DeadInst = NowDeadInsts.pop_back_val();
132 // This instruction is dead, zap it, in stages. Start by removing it from
134 SE.forgetValue(DeadInst);
136 for (unsigned op = 0, e = DeadInst->getNumOperands(); op != e; ++op) {
137 Value *Op = DeadInst->getOperand(op);
138 DeadInst->setOperand(op, 0);
140 // If this operand just became dead, add it to the NowDeadInsts list.
141 if (!Op->use_empty()) continue;
143 if (Instruction *OpI = dyn_cast<Instruction>(Op))
144 if (isInstructionTriviallyDead(OpI))
145 NowDeadInsts.push_back(OpI);
148 DeadInst->eraseFromParent();
150 } while (!NowDeadInsts.empty());
153 bool LoopIdiomRecognize::runOnLoop(Loop *L, LPPassManager &LPM) {
156 // The trip count of the loop must be analyzable.
157 SE = &getAnalysis<ScalarEvolution>();
158 if (!SE->hasLoopInvariantBackedgeTakenCount(L))
160 const SCEV *BECount = SE->getBackedgeTakenCount(L);
161 if (isa<SCEVCouldNotCompute>(BECount)) return false;
163 // If this loop executes exactly one time, then it should be peeled, not
164 // optimized by this pass.
165 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
166 if (BECst->getValue()->getValue() == 0)
169 // We require target data for now.
170 TD = getAnalysisIfAvailable<TargetData>();
171 if (TD == 0) return false;
173 DT = &getAnalysis<DominatorTree>();
174 LoopInfo &LI = getAnalysis<LoopInfo>();
176 SmallVector<BasicBlock*, 8> ExitBlocks;
177 CurLoop->getUniqueExitBlocks(ExitBlocks);
179 DEBUG(dbgs() << "loop-idiom Scanning: F["
180 << L->getHeader()->getParent()->getName()
181 << "] Loop %" << L->getHeader()->getName() << "\n");
183 bool MadeChange = false;
184 // Scan all the blocks in the loop that are not in subloops.
185 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
187 // Ignore blocks in subloops.
188 if (LI.getLoopFor(*BI) != CurLoop)
191 MadeChange |= runOnLoopBlock(*BI, BECount, ExitBlocks);
196 /// runOnLoopBlock - Process the specified block, which lives in a counted loop
197 /// with the specified backedge count. This block is known to be in the current
198 /// loop and not in any subloops.
199 bool LoopIdiomRecognize::runOnLoopBlock(BasicBlock *BB, const SCEV *BECount,
200 SmallVectorImpl<BasicBlock*> &ExitBlocks) {
201 // We can only promote stores in this block if they are unconditionally
202 // executed in the loop. For a block to be unconditionally executed, it has
203 // to dominate all the exit blocks of the loop. Verify this now.
204 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
205 if (!DT->dominates(BB, ExitBlocks[i]))
208 bool MadeChange = false;
209 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
210 // Look for store instructions, which may be memsets.
211 StoreInst *SI = dyn_cast<StoreInst>(I++);
212 if (SI == 0 || SI->isVolatile()) continue;
215 if (!processLoopStore(SI, BECount)) continue;
219 // If processing the store invalidated our iterator, start over from the
229 /// scanBlock - Look over a block to see if we can promote anything out of it.
230 bool LoopIdiomRecognize::processLoopStore(StoreInst *SI, const SCEV *BECount) {
231 Value *StoredVal = SI->getValueOperand();
232 Value *StorePtr = SI->getPointerOperand();
234 // Reject stores that are so large that they overflow an unsigned.
235 uint64_t SizeInBits = TD->getTypeSizeInBits(StoredVal->getType());
236 if ((SizeInBits & 7) || (SizeInBits >> 32) != 0)
239 // See if the pointer expression is an AddRec like {base,+,1} on the current
240 // loop, which indicates a strided store. If we have something else, it's a
241 // random store we can't handle.
242 const SCEVAddRecExpr *StoreEv =
243 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(StorePtr));
244 if (StoreEv == 0 || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
247 // Check to see if the stride matches the size of the store. If so, then we
248 // know that every byte is touched in the loop.
249 unsigned StoreSize = (unsigned)SizeInBits >> 3;
250 const SCEVConstant *Stride = dyn_cast<SCEVConstant>(StoreEv->getOperand(1));
252 // TODO: Could also handle negative stride here someday, that will require the
253 // validity check in mayLoopModRefLocation to be updated though.
254 if (Stride == 0 || StoreSize != Stride->getValue()->getValue())
257 // If the stored value is a byte-wise value (like i32 -1), then it may be
258 // turned into a memset of i8 -1, assuming that all the consequtive bytes
259 // are stored. A store of i32 0x01020304 can never be turned into a memset.
260 if (Value *SplatValue = isBytewiseValue(StoredVal))
261 if (processLoopStoreOfSplatValue(SI, StoreSize, SplatValue, StoreEv,
265 // If the stored value is a strided load in the same loop with the same stride
266 // this this may be transformable into a memcpy. This kicks in for stuff like
267 // for (i) A[i] = B[i];
268 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
269 const SCEVAddRecExpr *LoadEv =
270 dyn_cast<SCEVAddRecExpr>(SE->getSCEV(LI->getOperand(0)));
271 if (LoadEv && LoadEv->getLoop() == CurLoop && LoadEv->isAffine() &&
272 StoreEv->getOperand(1) == LoadEv->getOperand(1) && !LI->isVolatile())
273 if (processLoopStoreOfLoopLoad(SI, StoreSize, StoreEv, LoadEv, BECount))
276 //errs() << "UNHANDLED strided store: " << *StoreEv << " - " << *SI << "\n";
281 /// mayLoopAccessLocation - Return true if the specified loop might access the
282 /// specified pointer location, which is a loop-strided access. The 'Access'
283 /// argument specifies what the verboten forms of access are (read or write).
284 static bool mayLoopAccessLocation(Value *Ptr,AliasAnalysis::ModRefResult Access,
285 Loop *L, const SCEV *BECount,
286 unsigned StoreSize, AliasAnalysis &AA,
287 StoreInst *IgnoredStore) {
288 // Get the location that may be stored across the loop. Since the access is
289 // strided positively through memory, we say that the modified location starts
290 // at the pointer and has infinite size.
291 uint64_t AccessSize = AliasAnalysis::UnknownSize;
293 // If the loop iterates a fixed number of times, we can refine the access size
294 // to be exactly the size of the memset, which is (BECount+1)*StoreSize
295 if (const SCEVConstant *BECst = dyn_cast<SCEVConstant>(BECount))
296 AccessSize = (BECst->getValue()->getZExtValue()+1)*StoreSize;
298 // TODO: For this to be really effective, we have to dive into the pointer
299 // operand in the store. Store to &A[i] of 100 will always return may alias
300 // with store of &A[100], we need to StoreLoc to be "A" with size of 100,
301 // which will then no-alias a store to &A[100].
302 AliasAnalysis::Location StoreLoc(Ptr, AccessSize);
304 for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
306 for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I)
307 if (&*I != IgnoredStore &&
308 (AA.getModRefInfo(I, StoreLoc) & Access))
314 /// processLoopStoreOfSplatValue - We see a strided store of a memsetable value.
315 /// If we can transform this into a memset in the loop preheader, do so.
316 bool LoopIdiomRecognize::
317 processLoopStoreOfSplatValue(StoreInst *SI, unsigned StoreSize,
319 const SCEVAddRecExpr *Ev, const SCEV *BECount) {
320 // Verify that the stored value is loop invariant. If not, we can't promote
322 if (!CurLoop->isLoopInvariant(SplatValue))
325 // Okay, we have a strided store "p[i]" of a splattable value. We can turn
326 // this into a memset in the loop preheader now if we want. However, this
327 // would be unsafe to do if there is anything else in the loop that may read
328 // or write to the aliased location. Check for an alias.
329 if (mayLoopAccessLocation(SI->getPointerOperand(), AliasAnalysis::ModRef,
331 StoreSize, getAnalysis<AliasAnalysis>(), SI))
334 // Okay, everything looks good, insert the memset.
335 BasicBlock *Preheader = CurLoop->getLoopPreheader();
337 IRBuilder<> Builder(Preheader->getTerminator());
339 // The trip count of the loop and the base pointer of the addrec SCEV is
340 // guaranteed to be loop invariant, which means that it should dominate the
341 // header. Just insert code for it in the preheader.
342 SCEVExpander Expander(*SE);
344 unsigned AddrSpace = SI->getPointerAddressSpace();
346 Expander.expandCodeFor(Ev->getStart(), Builder.getInt8PtrTy(AddrSpace),
347 Preheader->getTerminator());
349 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
350 // pointer size if it isn't already.
351 const Type *IntPtr = TD->getIntPtrType(SI->getContext());
352 unsigned BESize = SE->getTypeSizeInBits(BECount->getType());
353 if (BESize < TD->getPointerSizeInBits())
354 BECount = SE->getZeroExtendExpr(BECount, IntPtr);
355 else if (BESize > TD->getPointerSizeInBits())
356 BECount = SE->getTruncateExpr(BECount, IntPtr);
358 const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1),
359 true, true /*nooverflow*/);
361 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
362 true, true /*nooverflow*/);
365 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
368 Builder.CreateMemSet(BasePtr, SplatValue, NumBytes, SI->getAlignment());
370 DEBUG(dbgs() << " Formed memset: " << *NewCall << "\n"
371 << " from store to: " << *Ev << " at: " << *SI << "\n");
374 // Okay, the memset has been formed. Zap the original store and anything that
376 DeleteDeadInstruction(SI, *SE);
381 /// processLoopStoreOfLoopLoad - We see a strided store whose value is a
382 /// same-strided load.
383 bool LoopIdiomRecognize::
384 processLoopStoreOfLoopLoad(StoreInst *SI, unsigned StoreSize,
385 const SCEVAddRecExpr *StoreEv,
386 const SCEVAddRecExpr *LoadEv,
387 const SCEV *BECount) {
388 LoadInst *LI = cast<LoadInst>(SI->getValueOperand());
390 // Okay, we have a strided store "p[i]" of a loaded value. We can turn
391 // this into a memcpy in the loop preheader now if we want. However, this
392 // would be unsafe to do if there is anything else in the loop that may read
393 // or write to the stored location (including the load feeding the stores).
394 // Check for an alias.
395 if (mayLoopAccessLocation(SI->getPointerOperand(), AliasAnalysis::ModRef,
396 CurLoop, BECount, StoreSize,
397 getAnalysis<AliasAnalysis>(), SI))
400 // For a memcpy, we have to make sure that the input array is not being
401 // mutated by the loop.
402 if (mayLoopAccessLocation(LI->getPointerOperand(), AliasAnalysis::Mod,
403 CurLoop, BECount, StoreSize,
404 getAnalysis<AliasAnalysis>(), SI))
407 // Okay, everything looks good, insert the memcpy.
408 BasicBlock *Preheader = CurLoop->getLoopPreheader();
410 IRBuilder<> Builder(Preheader->getTerminator());
412 // The trip count of the loop and the base pointer of the addrec SCEV is
413 // guaranteed to be loop invariant, which means that it should dominate the
414 // header. Just insert code for it in the preheader.
415 SCEVExpander Expander(*SE);
418 Expander.expandCodeFor(LoadEv->getStart(),
419 Builder.getInt8PtrTy(LI->getPointerAddressSpace()),
420 Preheader->getTerminator());
421 Value *StoreBasePtr =
422 Expander.expandCodeFor(StoreEv->getStart(),
423 Builder.getInt8PtrTy(SI->getPointerAddressSpace()),
424 Preheader->getTerminator());
426 // The # stored bytes is (BECount+1)*Size. Expand the trip count out to
427 // pointer size if it isn't already.
428 const Type *IntPtr = TD->getIntPtrType(SI->getContext());
429 unsigned BESize = SE->getTypeSizeInBits(BECount->getType());
430 if (BESize < TD->getPointerSizeInBits())
431 BECount = SE->getZeroExtendExpr(BECount, IntPtr);
432 else if (BESize > TD->getPointerSizeInBits())
433 BECount = SE->getTruncateExpr(BECount, IntPtr);
435 const SCEV *NumBytesS = SE->getAddExpr(BECount, SE->getConstant(IntPtr, 1),
436 true, true /*nooverflow*/);
438 NumBytesS = SE->getMulExpr(NumBytesS, SE->getConstant(IntPtr, StoreSize),
439 true, true /*nooverflow*/);
442 Expander.expandCodeFor(NumBytesS, IntPtr, Preheader->getTerminator());
445 Builder.CreateMemCpy(StoreBasePtr, LoadBasePtr, NumBytes,
446 std::min(SI->getAlignment(), LI->getAlignment()));
448 DEBUG(dbgs() << " Formed memcpy: " << *NewCall << "\n"
449 << " from load ptr=" << *LoadEv << " at: " << *LI << "\n"
450 << " from store ptr=" << *StoreEv << " at: " << *SI << "\n");
453 // Okay, the memset has been formed. Zap the original store and anything that
455 DeleteDeadInstruction(SI, *SE);